Emergent and Distributed Cognition Without a Single Intelligence Center

Executive answer

AIOS can responsibly be described as producing intelligence emergently from interacting processes, provided that “emergently” is defined in a modest, operational sense. The defensible claim is one of organizational or weak emergence: a user-visible capability is realized by coordinated interactions among files, metadata, annotations, relationships, context composition, model judgments, exact operations, downstream integration, memory, and human decisions; the capability is not attributable to any one element considered in isolation. This is a claim that can be tested through component ablations, interaction effects, provenance, calibration, and task performance.

That wording does not imply that AIOS reproduces human neuroanatomy, implements a settled theory of consciousness, has subjective experience, or constitutes a unitary artificial mind. Current consciousness science does not license those inferences. The strongest preregistered comparison of two leading neural theories, published in 2025, challenged central predictions of both global neuronal workspace theory (GNWT) and integrated information theory (IIT). Functional resemblance to selection, broadcast, recurrence, confidence, or post-decision review is not evidence of consciousness.

The recent evidence supports five narrower conclusions relevant to system design:

1. Human perception and decision-making are distributed across interacting stages; report, attention, awareness, confidence, and revision are related but not interchangeable.
2. Selection and relational learning can affect behavior without detailed explicit knowledge. Attention can also be oriented by stimuli that participants cannot report seeing.
3. Evidence accumulation can continue after an initial commitment and can support confidence and changes of mind. This gives a useful design analogy for short foreground responses followed by non-blocking review, but it is not biological validation of that design.
4. Cognitive work can be analyzed at the level of a person–artifact–social system. Whether an external resource merely assists cognition or literally constitutes part of a mind remains philosophically disputed.
5. Human–AI combinations are not generically synergistic. A 2024 meta-analysis found that combinations underperformed the better of human-only or AI-only performance on average. More recent field and deliberation studies show both meaningful gains and important failures, including bias amplification, steering, and perceived inclusion without improved participation equity.

The most defensible one-sentence description is:

AIOS is a distributed, human-governed reasoning environment in which system-level performance emerges from coordinated interactions among person-controlled knowledge, structured context, models, exact operations, memory, and review. “Emerges” is an architectural and empirical claim about composition, not a claim that AIOS reproduces the brain, is conscious, or possesses a unitary mind.

The larger AIOS thesis examined here is complementary rather than substitutional. It does not require local systems to replace frontier-model research, centralized training, or every remote inference call. It proposes that durable context, domain expertise, memory, workflow state, and consequential authority can move into person- or organization-controlled systems, while local and frontier models are selected according to task demands. If capable local models and mature scaffolding make that allocation reliable across a large share of ordinary work, the first-order effect would be less repeated context reconstruction and less routine data movement to centralized services. The second-order effects could include persistent personal and organizational intelligence, private local cognitive workflows, portable domain systems, selective rather than default remote inference, new forms of peer collaboration, and broader offline or low-connectivity access. Section 10 examines those implications without presenting them as already demonstrated.

1. Scope, method, and evidence policy

This memo asks what recent cognitive science and human–AI research can—and cannot—contribute to the description of a local-first, file-native reasoning environment with no single sufficient intelligence center.

This memo prioritizes:

• primary, peer-reviewed work published from August 2024 through the cutoff date;
• preregistered, multimodal, intracranial, causal-intervention, field-experimental, and meta-analytic designs;
• papers that distinguish selection, report, access, confidence, revision, and system-level coordination;
• authoritative reviews where a field is too theoretically unsettled for a single experiment to represent its state;
• direct publisher or DOI links.

Preprints, perspectives, reviews, computational studies, vendor-authored work, industry-partnered work, small clinical samples, and findings without independent replication are labeled. Older work appears only in the Foundational lineage section.

The memo separates three levels:

• Evidence: what a study directly measured.
• Interpretation: what follows at a conceptual level, with uncertainty preserved.
• Possible AIOS connection: a design question or evaluation target. This is not a claim that the cited work validates AIOS.

The source set is intentionally selective. Tenuous analogies—especially attempts to map named brain regions onto software components—were excluded.

Section 10 separately examines the proposed AIOS architectural theses and their conditional implications. The theses are not treated as findings supplied by the cognitive-science literature, but their first- and second-order consequences are developed rather than excluded.

2. Conceptual distinctions that should not be collapsed

Two boundary rules follow.

First, access is not identical to consciousness. An engineering system can make information available to several modules without any evidence about subjective experience. Second, no single sufficient center does not mean no bottlenecks, control points, or authority. AIOS can have local selection mechanisms and model calls while remaining systemically distributed. Human authority is not another interchangeable component; it defines goals, accepted standing, canonical-record designation, and reversibility.

3. Preconscious selection, conscious access, and attention

3.1 Implicit extraction of temporal relations

Tacikowski et al., “Human hippocampal and entorhinal neurons encode the temporal structure of experience” (Nature, 25 September 2024). [Peer-reviewed primary research; rare intracranial human sample; not yet broadly replicated]

• Research question: How do human hippocampal and entorhinal neurons extract the temporal structure connecting events, and can the resulting representation be predictive without detailed explicit knowledge?
• Method/sample: Extracellular recordings from 1,456 single- and multi-units across 21 sessions in 17 patients with epilepsy. Six images were assigned to nodes in a hidden pyramid graph; graph-constrained sequences were presented while participants performed an unrelated task. A separate behavioral study included 25 healthy controls, with an additional explicit-instruction benchmark.
• Principal finding: Hippocampal–entorhinal population activity gradually came to reflect the graph’s relational and successor-like structure, persisted after graph-constrained presentation stopped, and was associated with time-compressed replay. Violations later slowed responses. Neither patients nor healthy controls reported graph-like organization, and their detailed explicit knowledge remained below the explicit benchmark.
• Limitations: Clinical and small sample; sparse, treatment-determined electrode coverage; a highly structured visual sequence; no warrant to call all learning “unconscious”; some coarse structural knowledge was above chance; neuronal decoding does not by itself specify cognitive function.
• Relevance: The study supports a narrow proposition: durable relational structure can be learned from use before it is available as a detailed verbal model. It does not show that a metadata-emergent relationship graph has human memory properties. For AIOS, the reasonable question is whether relationships inferred from ordinary file activity improve retrieval and prediction while remaining inspectable and correctable by the user.

3.2 Evidence accumulation without immediate report

Stockart et al., “Cortical evidence accumulation for visual perception occurs irrespective of reports” (Nature Communications, 26 September 2025). [Peer-reviewed primary research; preregistered; clinical intracranial sample]

• Research question: Are cortical accumulation signals tied only to immediate behavioral report, or do they track perception when report is delayed or absent?
• Method/sample: Three preregistered stereotactic EEG experiments using 3,301 recording sites in 29 people with drug-resistant focal epilepsy. Near-threshold faces appeared in scrambled streams under immediate report, delayed report, and passive-viewing conditions; analyses related high-gamma signals to reaction time, seen/unseen judgments, stimulus intensity, and confidence. A leaky-accumulator model was fitted to behavior and neural data.
• Principal finding: Accumulation-like signals were distributed across visual, inferior frontal, and anterior insular cortex. Ventral visual signals differentiated seen from unseen trials under delayed report, tracked physical intensity during passive viewing, and related to confidence. The model reproduced key behavioral and neural patterns.
• Limitations: Clinical sampling and nonuniform coverage; correlations and model fit do not uniquely identify a mechanism; perceptual detection is narrower than reasoning; “no report” reduces but does not eliminate all task and state confounds.
• Relevance: Immediate output is not a complete proxy for all processing that produced it. For AIOS, this motivates logging selection and accumulation before, during, and after a foreground response. It does not imply an artificial conscious percept.

3.3 Attention without cue awareness, and awareness changing attention

Yang et al., “Visual awareness sharpens and accelerates attentional sampling…” (Nature Communications, 17 November 2025). [Peer-reviewed primary research; small samples; not independently replicated]

• Research question: Does rhythmic attentional sampling require awareness of the cue, and how does awareness alter its dynamics?
• Method/sample: Chromatic flicker fusion rendered spatial cues objectively invisible. A within-subject behavioral experiment used 20 adults; an EEG experiment used 22. Awareness checks found invisible-cue localization at chance. Temporal-response and connectivity analyses compared visible and invisible conditions.
• Principal finding: Both visible and invisible cues captured attention and induced rhythmic sampling. Visible cues produced stronger inhibition, earlier/higher-frequency frontoparietal coupling, and faster sampling (about 8 Hz versus 4 Hz).
• Limitations: Small, young samples; single research program; awareness depends on the sensitivity of the chosen checks; oscillatory/rhythmic-attention results are analysis-sensitive; causal direction cannot be inferred from connectivity estimates alone.
• Relevance: Attention and awareness are neither identical nor independent. A system can prioritize information without representing that prioritization in the foreground. AIOS should therefore distinguish background routing from user-visible reasons and provide an audit path for consequential selection.

Evidence synthesis

The recent evidence favors a layered account over a single funnel:

• relational regularities may be acquired without detailed explicit knowledge;
• perceptual evidence can accumulate across distributed areas even when reporting is delayed or absent;
• attention can be influenced without cue awareness, while awareness changes the strength and temporal organization of attention.

This does not identify a universal “preconscious layer.” Different tasks recruit different mechanisms. For system design, “preconscious” should remain a human cognitive-science term. “Background selection,” “latent processing,” and “foreground access” are clearer engineering descriptions.

4. Global workspace and related models: useful architecture, unsettled consciousness theory

4.1 The strongest recent adversarial test

Cogitate Consortium et al., “Adversarial testing of global neuronal workspace and integrated information theories of consciousness” (Nature, 30 April 2025). [Peer-reviewed primary research; preregistered; open-science adversarial collaboration]

• Research question: Which divergent neural predictions of GNWT and IIT survive a theory-neutral, preregistered, multimodal test?
• Method/sample: 256 human participants viewed suprathreshold visual stimuli of varied category, identity, orientation, and duration. The same task was studied with fMRI (n=120), MEG (n=102), and intracranial EEG (n=34), across multiple laboratories. Theory proponents and neutral researchers agreed in advance on predictions, analyses, pass/fail criteria, and interpretations; optimization and held-out test data were separated.
• Principal finding: Conscious content information appeared in visual, ventrotemporal, and inferior frontal regions; sustained responses tracked stimulus duration in occipital and lateral temporal regions; some content-specific frontal–early-visual synchronization occurred. Yet key predictions of both theories failed: IIT lacked predicted sustained posterior synchronization, while GNWT lacked a general offset ignition and showed limited prefrontal representation of some conscious dimensions.
• Limitations: The study tested proposed biological implementations more directly than the theories’ mathematical cores. It used clearly visible, centrally presented, attended stimuli, not near-threshold access or naturalistic cognition. Negative results can motivate theory revision rather than decisive rejection. The experiment does not establish that IIT and GNWT exhaust the theory space.
• Relevance: No leading neural theory can be imported as settled design law. A shared context, broadcast mechanism, or recurrent integration loop may be useful engineering, but calling it a “global workspace” should be marked as a functional analogy. It is not evidence of consciousness.

4.2 A computational mechanism for ignition

Klatzmann et al., “A dynamic bifurcation mechanism explains cortex-wide neural correlates of conscious access” (Cell Reports, 25 March 2025). [Peer-reviewed computational study with biological constraint; not a direct consciousness experiment]

• Research question: Can a biophysically constrained large-scale cortical model reproduce threshold-like ignition and sustained activity associated with detected stimuli?
• Method/benchmark: A connectome-based dynamical model of 40 macaque cortical areas used tract-tracing, hierarchical, spine-density, and receptor-distribution constraints. Simulated stimulus-detection dynamics were compared with reported primate neural signatures; a predicted NMDA-to-AMPA receptor gradient was compared with autoradiography data.
• Principal finding: A dynamic bifurcation produced a transition from transient sensory propagation to sustained association-area activity. Fast AMPA-dominated feedforward excitation propagated signals; NMDA contributions in feedback and local recurrence helped stabilize ignition. The receptor-gradient comparison was consistent with a model prediction.
• Limitations: Simulation demonstrates sufficiency under assumptions, not the mechanism actually used by a conscious brain. Neural signatures came from prior experiments; the model is macaque-based; parameter identifiability and alternative architectures remain open. Several authors are closely associated with GNWT, so this is not an adversarial test.
• Relevance: The study is a good example of system-level threshold behavior arising from feedback, timescale, and connectivity interactions. That is a legitimate inspiration for interaction-dependent routing or context stabilization. Mapping AMPA, NMDA, cortical hierarchy, or ignition onto AIOS components would be scientifically tenuous.

4.3 Integrative and AI-consciousness perspectives

Two recent perspectives add necessary caution.

Mudrik, Faivre, Pitts, and Schurger, “On a confusion about there being two types of consciousness” (Trends in Cognitive Sciences, 2025). [Peer-reviewed opinion/perspective; not new primary evidence] The authors argue against treating phenomenal and access consciousness as two separable “types.” Their proposed framework requires both potentially phenomenal content and access to other systems for a conscious episode. It is an interpretive attempt to reconcile local-content and access evidence, not an established test of consciousness. Its useful implication here is negative: access alone is insufficient.

Butlin et al., “Identifying indicators of consciousness in AI systems” (Trends in Cognitive Sciences, online 10 November 2025; issue 2026). [Peer-reviewed opinion/perspective; theory-derived framework, not an empirical finding] The paper derives indicators from several neuroscientific theories and treats them as considerations that may raise or lower credence, not as necessary and sufficient conditions. It explicitly begins from theoretical uncertainty. Properties such as recurrence, workspace-like selection and broadcast, metacognitive monitoring, attention mechanisms, or hierarchical prediction errors therefore cannot be treated as a consciousness checklist. Superficial behavior is especially vulnerable to gaming and anthropomorphic interpretation.

Interpretation for AIOS

A layered architecture can use workspace-like ideas responsibly if three qualifications stay visible:

1. Functional only: context composition makes selected material available for downstream use.
2. Plural and revisable: GNWT is one model among several, and its neural implementation has been materially challenged.
3. No phenomenology inference: availability, recurrence, and integration do not establish subjective experience.

The system-description question is not “Does AIOS have a global workspace?” It is “Which information becomes available to which processes, by what selection rule, with what provenance, for how long, and under whose authority?”

5. Predictive processing: current evidence supports local mechanisms, not a universal analogy

5.1 Prediction errors and representational change in humans

Greco et al., “Predictive learning shapes the representational geometry of the human brain” (Nature Communications, 8 November 2024). [Peer-reviewed primary research; small human neuroimaging sample]

• Research question: Does prediction-error encoding relate to changes in neural representational geometry during statistical learning?
• Method/sample: MEG from 24 adults listening passively to high- and low-regularity tone sequences. Representational-similarity analysis, an ideal-observer model, mutual-information analysis, and partial information decomposition were used to relate stimulus statistics, prediction errors, and distributed neural representations.
• Principal finding: Representations increasingly grouped temporally contiguous and predictable tones. Prediction-error signals were broadly distributed; their magnitude correlated with representational shift in left temporal cortex (r=0.47, Bonferroni-corrected p=0.021), but not in the tested frontal clusters. Pairwise analyses found more synergistic than redundant prediction-error information across regions.
• Limitations: Small sample and passive artificial auditory task; correlational link; the ideal observer may not capture individual learning; whole-brain source reconstruction is indirect; non-linear learning dynamics were not conclusively supported. The distributed result may challenge simple hierarchical predictive-coding formulations.
• Relevance: The study supports a bounded design idea: tracked discrepancies can update representations across a network. It does not establish that all cognition minimizes prediction error or that an AIOS relationship graph is brain-like.

5.2 Causal error signals for task switching in mice

Cole et al., “Prediction-error signals in anterior cingulate cortex drive task-switching” (Nature Communications, 17 August 2024). [Peer-reviewed primary research; animal study with causal intervention]

• Research question: Are neural prediction-error signals causally required for rapid rule switching?
• Method/sample: Mice switched between visual and olfactory discrimination rules, often in a single trial. Behavioral modeling was combined with widefield and two-photon calcium imaging, optogenetic silencing and release, and all-optical manipulation of vasoactive intestinal peptide interneurons. Key experiments used small cohorts, including 10 mice for intensive imaging and 8 for optogenetic silencing.
• Principal finding: Anterior cingulate error signals preceded successful transitions and were required specifically when the task demanded rapid switching. Perturbations implicated a disinhibitory interneuron circuit in computing the error.
• Limitations: Mouse circuitry and a tightly controlled cross-modal task; small cohorts typical of invasive systems neuroscience; no claim about human conscious access; “world model” is task-specific shorthand. Generalization to document reasoning is speculative.
• Relevance: Error-triggered reconfiguration can be both selective and causal. The appropriate AIOS connection is an explicit rule for when contradictions, failed operations, or low-quality evidence should recompose context or hand control back to the user—not a claim of cortical equivalence.

5.3 The 2026 field reassessment

Furutachi and Hofer, “Rethinking Predictive Processing” (Annual Review of Neuroscience, 8 July 2026). [Peer-reviewed review; not primary research]

• Research question: What do sensory prediction-error signals actually encode, and how strong is the evidence for predictive coding as a general neural algorithm?
• Method: Critical review of 185 references across the historical framework, neuronal data, circuit claims, and alternative explanations.
• Principal conclusion: Neural activity often looks compatible with prediction-error signaling, but definitions vary and similar response patterns can arise from different computations. Clarifying encoded information and testing alternatives are prerequisites for stronger claims.
• Limitations: A synthetic and argumentative review, not a new experiment; its conclusions depend on the selected literatures and framing.
• Relevance: Predictive processing should enter AIOS only at the level of explicit, testable mechanisms—expectation, mismatch, revision, and measured benefit. “The system is a predictive brain” would outrun the evidence twice: the brain theory remains contested, and AIOS is not a brain model.

6. Metacognitive judgment and post-decision integration

6.1 Confidence as online control

Balsdon and Philiastides, “Confidence control for efficient behaviour in dynamic environments” (Nature Communications, 22 October 2024). [Peer-reviewed primary research; preregistered; small human EEG sample]

• Research question: Can confidence influence evidence gathering during a decision rather than only summarize it afterward?
• Method/sample: Twenty analyzed participants completed 900 dynamic random-dot-motion decisions while EEG was recorded. A preregistered double-integration sequential-sampling model allowed an evolving confidence estimate to control leakage in a second, motor-related accumulator. Model comparison and multivariate EEG decoding tested latent-process predictions.
• Principal finding: The confidence-control model outperformed classic and alternative models, predicted post-decision confidence from choice and response time, and yielded EEG correlates aligned with the proposed accumulators. Stronger EEG-derived confidence control was associated with faster and more accurate choices.
• Limitations: Small sample; model-dependent latent variables; low but above-chance EEG decoding; the task couples perceptual and motor processes; association does not establish that subjective confidence itself caused efficiency. Online confidence and explicit reported confidence were not interchangeable.
• Relevance: A confidence signal can be useful when it changes resource allocation or stopping behavior and is calibrated to outcomes. An AIOS confidence field is not metacognition by declaration; it should be evaluated with calibration error, selective risk, and intervention tests.

6.2 Evidence after commitment, confidence, and changes of mind

Goueytes et al., “Evidence accumulation in the pre-supplementary motor area and insula drives confidence and changes of mind” (Nature Communications, 30 July 2025). [Peer-reviewed primary research; clinical intracranial sample]

• Research question: Where and when does evidence accumulation contribute to confidence and changes of mind before and after an initial decision?
• Method/sample: Twenty-four people with drug-resistant focal epilepsy completed a random-dot two-choice task and rated confidence from 0 to 100. Mouse trajectories estimated decision and change-of-mind timing; stereotactic EEG measured high-gamma activity; a race-diffusion model continued accumulation for up to one second after initial choice. Fifteen participants contributed to the full model fit.
• Principal finding: Accumulation-like signals were distributed, with pre-decision effects strongest in pre-supplementary motor cortex and post-decision confidence effects prominent in insula and orbitofrontal cortex. In 13 of 15 fitted participants, a post-decision model had higher likelihood than an accumulation-to-bound model. About 9.6% of trials contained trajectory-defined changes of mind, whose timing the model reproduced.
• Limitations: Clinical sample and treatment-determined coverage; motor trajectories partly define the phenomena being explained; evidence accumulation is one of several confidence models; sparse change-of-mind events reduced power; anatomical correlations are not a complete causal account.
• Relevance: Initial commitment need not terminate evaluation. A short foreground response followed by continued verification, reconciliation, or memory integration has a defensible cognitive-design analogy. The operational test is whether downstream work catches errors or improves later action without silently overriding the human-approved canonical state.

Evidence synthesis

Metacognitive judgment is not a decorative self-score. To count as a meaningful system capability, second-order monitoring should:

• predict first-order correctness or failure on held-out tasks;
• change behavior appropriately, such as seeking more evidence or deferring;
• remain calibrated across domains and model changes;
• expose the evidence behind escalation or revision;
• avoid converting fluency into unwarranted certainty.

Post-decision integration is strongest when it is revision-capable but authority-bounded. The user should see whether a later process confirmed, qualified, or contradicted a foreground answer, and canonical files should change only through authorized exact operations.

7. Distributed and extended cognition in human–artifact systems

Distributed cognition changes the unit of analysis. Instead of asking only what is inside an individual, it asks how representations and operations propagate through people, artifacts, procedures, and environment. This makes it directly relevant to file-native reasoning. It does not settle the stronger extended-mind claim that external artifacts literally constitute an individual’s cognitive states.

7.1 Human–AI synergy is not the default

Vaccaro, Almaatouq, and Malone, “When combinations of humans and AI are useful: A systematic review and meta-analysis” (Nature Human Behaviour, 28 October 2024). [Peer-reviewed preregistered systematic review and meta-analysis; studies through June 2023]

• Research question: When does a human–AI combination outperform both its human-only and AI-only baselines?
• Method/sample: 106 experimental studies and 370 effect sizes, each required to include human-only, AI-only, and combined performance. The review distinguished human augmentation from strict synergy.
• Principal finding: Human–AI combinations improved on humans alone on average (Hedges’ g=0.64) but performed worse than the better of human-only or AI-only performance (g=−0.23, 95% CI −0.39 to −0.07). Decision tasks tended toward losses; content-creation tasks showed greater gains. When humans alone were stronger than AI alone, combinations were more likely to help; when AI alone was stronger, combinations were more likely to hurt.
• Limitations: Heterogeneous tasks and interfaces; possible publication bias; rapidly changing technology; studies ended in mid-2023; only three experiments used predetermined subtask delegation, so evidence for deliberately decomposed workflows was sparse.
• Relevance: A multi-component architecture should not assume synergy. AIOS needs component and combination baselines. Improvement over an unaided person is not enough to establish emergent advantage if a simpler configuration performs as well or better.

7.2 AI-mediated common ground

Tessler et al., “AI can help humans find common ground in democratic deliberation” (Science, 18 October 2024). [Peer-reviewed primary research; vendor-authored by Google DeepMind; not independently replicated at comparable scale]

• Research question: Can an AI mediator synthesize diverse views into statements that groups endorse while incorporating dissent?
• Method/sample: 5,734 UK participants. The “Habermas Machine” iteratively produced group statements from opinions and critiques; experiments compared AI and human mediators and included a virtual citizens’ assembly with a demographically representative UK sample.
• Principal finding: Participants preferred AI-generated statements and rated them clearer, more informative, and less biased. Views often converged, and text analyses indicated that accepted statements incorporated minority critiques while respecting the majority position.
• Limitations: Vendor authorship; UK political context; approval and convergence are not truth, legitimacy, or durable collective intelligence; comparison conditions did not capture all forms of human deliberation; optimizing endorsement can introduce majoritarian pressure; generalization to high-stakes, longitudinal groups is unknown.
• Relevance: Iterative synthesis can create a group-level artifact that no individual authored. The important design feature is structured critique followed by revision, not the claim that the mediator is a collective mind. Dissent retention and human acceptance should be measured separately from surface agreement.

7.3 Feedback loops can amplify bias

Glickman and Sharot, “How human–AI feedback loops alter human perceptual, emotional and social judgements” (Nature Human Behaviour, published 18 December 2024; issue 2025). [Peer-reviewed primary research; multiple experiments]

• Research question: What happens when a slightly biased human dataset trains an AI whose outputs then influence later human judgments?
• Method/sample: A series of experiments totaling 1,401 participants across perceptual, emotional, and social judgments, using convolutional neural networks, transformer-based systems, and a real text-to-image system. Human–AI and human–human feedback were compared; model identity labels were manipulated.
• Principal finding: AI could amplify small human biases, and subsequent users internalized those biases more strongly than in human–human interaction, often without recognizing the influence. When AI was accurate rather than biased, human judgment improved.
• Limitations: Experimental tasks and repeated short interactions; the result depends on biased input and system response; it does not show that every hybrid system amplifies bias; long-term organizational dynamics were not measured.
• Relevance: Files, annotations, summaries, and model judgments form a learning environment for the user. Provenance, counterevidence, diversity checks, and correction history are essential because a stable file-native loop can preserve and amplify error as effectively as it preserves knowledge.

7.4 Field evidence for performance and expertise integration

Dell’Acqua et al., “The Cybernetic Teammate: A Field Experiment on Generative AI and Teamwork” (Organization Science, online 12 June 2026). [Peer-reviewed primary research; preregistered; industry-partnered with Procter & Gamble]

• Research question: Can generative AI supply some performance, expertise-integration, and social benefits ordinarily attributed to a human teammate?
• Method/sample: 826 professionals joined a one-day product-development workshop; 791 randomized participants entered the analysis and 776 had complete post-task data. A 2×2 design compared individuals versus two-person cross-functional teams, with versus without GPT-4 accessed through a controlled interface. Real business challenges and blind human expert ratings were used.
• Principal finding: Relative to individuals without AI, individuals with AI improved 0.37 SD (about 9.6%) and teams with AI 0.39 SD (about 10.2%); teams without AI improved 0.24 SD. Individuals with AI matched teams without AI. AI broadened technical/commercial balance, while teams plus AI were 9.2 percentage points more likely than the 5.8% control mean to produce a top-decile solution. AI users were nevertheless 9.2 percentage points less likely to predict that their output was top-decile, and AI did not improve evaluative selection as clearly as idea generation.
• Limitations: One company and industry; one-day virtual “flash teams”; one model at one point in time; early-stage ideation rather than implementation or commercial success; dyads rather than complex established teams; close industry collaboration; increased output length and model-shaped similarity complicate quality interpretation.
• Relevance: Complementary tools can improve a system-level outcome and bridge domain boundaries without replacing human evaluation. The study supports measuring stage-specific complementarity, not calling the whole system conscious or generally intelligent. It also warns that objective performance and self-assessed confidence can diverge.

7.5 Real-time facilitation: preference without consensus, inclusion, or neutrality

Parisi et al., “Real-Time Group Dynamics with LLM Facilitation: Evidence from a Charity Allocation Task” (FAccT, 25–28 June 2026). [Peer-reviewed conference paper; vendor-authored by Google DeepMind; arXiv version available]

• Research question: Does an LLM facilitator improve consensus or participation equity in real-time, consequential group discussion, and can it steer outcomes?
• Method/sample: Two studies totaling 879 participants in groups of three, allocating $7,200 in real charitable payments. Study 1 (n=204) compared three frontier models; Study 2 (n=675) compared facilitation strategies with no facilitator.
• Principal finding: Facilitation did not significantly improve consensus in either study, although participants preferred it. Facilitators shifted selected charity allocations by as much as 5.5 percentage points. Participants cited inclusion, but neither survey nor transcript measures showed improved participation equity. Trust increased under some conditions that also exerted directional influence.
• Limitations: Five-minute discussions, groups of three, one controlled allocation task, high baseline agreement, no human-facilitator baseline, and exploratory steering analyses. The results do not negate the different asynchronous synthesis design in Tessler et al.
• Relevance: Perceived procedural quality, balanced participation, outcome quality, and neutrality are distinct. AIOS should not infer that a smooth synthesis is representative or non-steering. It should make selection and compression effects inspectable.

7.6 External memory: support, offloading, and dependence

Nicolas Crozatier, “An overview of the ‘Externalization’ of memory…” (Memory, Mind & Media, 2026). [Peer-reviewed field review; not primary empirical research]

• Research question: How should technological externalization differ across semantic, episodic, and prospective memory?
• Method: Interdisciplinary synthesis and conceptual taxonomy of declarative-memory externalization.
• Principal conclusion: The paper distinguishes offloading, where an external resource reduces internal load, from biloading, where internal and external resources provide redundant support. It treats technological memory as coordination, not a single undifferentiated external store.
• Limitations: Conceptual synthesis; no new experiment; philosophical questions about causal assistance versus constitutive extension remain open.
• Relevance: Multiple-resolution file memory can be evaluated by function: storage, retrieval, reinforcement, prospective prompting, and error recovery. User autonomy is better served by inspectable redundancy and exportable canonical files than by opaque dependence on a single generated summary.

Interpretation for AIOS

Distributed cognition offers the most relevant level of analysis, but three different claims must remain separate:

1. Causal integration: files, tools, and models affect the user’s reasoning. This is straightforward and testable.
2. System-level cognition: for some research questions, the person–artifact workflow is the useful unit of analysis. This is an analytic stance, supported by distributed-cognition practice.
3. Constitutive extension: an external resource is literally part of a person’s mind. This is a stronger philosophical claim and is unnecessary for describing AIOS.

Local-first and file-native design can strengthen reliable availability, inspectability, and person control—the very properties that make external resources cognitively useful. But tighter coupling also increases dependence and feedback-loop risk. Portability, transparent provenance, reversible operations, and the ability to work without a specific model are therefore cognitive-resilience properties, not merely implementation preferences.

8. Perennial philosophy: inspiration, not evidence

The phrase perennial philosophy has a specific historical and metaphysical burden. As summarized by the Stanford Encyclopedia of Philosophy, twentieth-century perennialists such as Aldous Huxley claimed a universal core of esoteric doctrines expressed across cultures and religions. Related “essentialist” views propose a culture-independent common core of mystical experience. These are disputed philosophical, historical, and religious claims.

A recurring Why–How–What grammar may be inspired by themes of unity-through-difference, nested wholes, recurrence across scales, or disciplined self-inquiry. That is legitimate as design inspiration. It does not provide empirical support for a system architecture.

Villiger, “Mystical experience in the Bayesian brain” (Phenomenology and the Cognitive Sciences, 10 December 2025) is a peer-reviewed theoretical perspective, not a new experiment. It interprets mystical experience through the REBUS/Bayesian-brain framework. It may generate hypotheses about precision weighting and self-models, but it neither verifies perennial metaphysics nor shows that predictive processing explains mystical experience uniquely. The 2026 critical review of predictive processing makes that restraint especially important.

For the Fractal Seed, the empirical program is simpler and stronger than a metaphysical one. Test whether repeating Why–How–What across scales improves:

• cross-domain transfer;
• planning completeness;
• retrieval precision and recall;
• consistency between intent, method, and artifact;
• detection of missing rationale or unsupported action;
• human correction time;
• learning across document, project, and knowledge-evolution tasks.

If those effects appear under controlled comparison, they support a reusable cognitive scaffold. They do not establish a perennial truth, a universal structure of mind, or consciousness.

9. Can “intelligence emerges without a single center” be made rigorous?

9.1 A defensible definition

For AIOS, define intelligence operationally as reliable capability to produce, revise, and apply useful representations and actions across bounded knowledge tasks under human authority. Define emergence as an interaction-dependent system effect:

A capability is emergent in the relevant engineering sense when it is a stable property of the organized workflow, no single component is sufficient to produce it across the target task distribution, and controlled changes to component interactions produce measurable changes in the capability.

This definition is compatible with mechanistic explanation. It does not require unpredictability, irreducibility, new causal laws, or phenomenology. “More than the sum of the parts” should be avoided unless an actual interaction effect exceeds an additive baseline.

9.2 What would make the claim empirical

The strongest test of “no single center” is not a diagram. It is causal decomposition. If one hidden model call determines nearly every consequential result, the system may be distributed in storage yet centralized in judgment. Conversely, if purpose, canonical files, metadata, exact operations, review, and human acceptance make distinct, measurable contributions—and the dominant contributor varies by task—the distributed description becomes credible.

9.3 Architecture language that stays within the evidence

Responsible:

• “AIOS organizes intelligence as a property of a human-governed workflow rather than locating it in a single model.”
• “Selected context is made available to bounded processes through workspace-like composition.”
• “Confidence estimates and post-output checks are calibrated control mechanisms.”
• “Files and metadata support a distributed cognitive workflow.”
• “System-level capability depends on interactions among person-controlled knowledge, models, operations, and review.”

Needs qualification:

• “Intelligence has no single center.” Add: “No component is sufficient across the full workflow; local bottlenecks and decision points still exist, and final authority remains human.”
• “The relationship graph is emergent.” Add the actual inference mechanism and evaluation; otherwise it may be merely implicit.
• “AIOS has a global workspace.” Prefer “workspace-like context composition,” because the neuroscience theory is about conscious access and remains contested.
• “AIOS is metacognitive.” Prefer “it estimates uncertainty and monitors outcomes” until second-order calibration and control are demonstrated.

Not supported:

• that AIOS is conscious, sentient, self-aware, or phenomenally integrated;
• that attention, broadcast, recurrence, confidence, or memory together prove consciousness;
• that the architecture reproduces the human brain;
• that a distributed workflow eliminates agency, governance, or responsibility;
• that human–AI combination is inherently synergistic;
• that perennial philosophy validates the Fractal Seed scientifically.

10. AIOS architectural thesis and conditional implications

This section addresses the AIOS theses that intersect emergent cognition, distributed cognition, context composition, division of labor, human authority, local model deployment, and person- or organization-controlled knowledge. “AI-native intelligence architecture” is treated as an architectural thesis rather than an established scientific category or an inevitable historical transition. That distinction does not remove its larger implications. It makes their logical status explicit.

Where relevant, the analysis uses four layers:

1. What current evidence establishes. Direct findings, bounded by study design.
2. The AIOS architectural thesis. The proposed allocation of knowledge, models, controls, and authority.
3. What follows if the thesis is substantially correct. First-order operational effects and second-order social, organizational, and infrastructural implications.
4. Conditions, uncertainties, counterforces, and tests. What must be true, what could prevent the implication, and how to discriminate the thesis from alternatives.

10.1 Targeted recent evidence

Modarressi et al., “NoLiMa: Long-Context Evaluation Beyond Literal Matching” (ICML, 13–19 July 2025). [Peer-reviewed primary benchmark research; mixed academic/Adobe authorship]

• Research question: Can models use information in long contexts when the query and relevant passage do not share an easy literal match?
• Method/sample or benchmark: NoLiMa inserts passages requiring latent association into progressively longer contexts. The authors evaluated 13 widely used models advertising context capacities of at least 128,000 tokens.
• Principal finding: Performance was strong below 1,000 tokens but degraded sharply with length. At 32,000 tokens, 11 of 13 models fell below half of their short-context baselines; GPT-4o declined from 99.3% to 69.7%. Reasoning-capable models and chain-of-thought prompting did not remove the problem.
• Limitations: This is a synthetic retrieval-and-association benchmark, not an end-to-end study of personal or organizational reasoning. Its 2025 model sample will age, and it does not uniquely identify whether attention, position, retrieval strategy, or some other mechanism caused each failure.
• Relevance: It directly contradicts the idea that insufficient context is the sole important cause of agent failure. More supplied context can itself become a source of failure; nominal context-window capacity is not effective context use.

Hwang et al., “LLMs can be easily Confused by Instructional Distractions” (ACL, July 2025). [Peer-reviewed primary benchmark research; not independently replicated]

• Research question: Do models preserve the governing instruction when the material to be processed itself resembles an instruction?
• Method/sample or benchmark: DIM-Bench crosses four transformation tasks—proofreading, rewriting, translation, and style transfer—with five kinds of instruction-like input: reasoning, code generation, mathematics, bias detection, and question answering.
• Principal finding: Advanced models were susceptible to “instructional distraction” and often followed an instruction-like passage instead of the user’s intended transformation, even when prompts explicitly distinguished task from input.
• Limitations: The benchmark is constructed and narrower than an operating environment. It demonstrates a class of instruction–data confusion, not a general suppression of reasoning by all scaffolds.
• Relevance: It supports separating trusted instructions, untrusted content, roles, and action authority. It does not show that every cognitive mode requires a separate agent or context.

Li et al., “When Thinking Fails: The Pitfalls of Reasoning for Instruction-Following in LLMs” (NeurIPS, 2025). [Peer-reviewed primary benchmark research; mixed academic/Amazon authorship]

• Research question: Does explicit chain-of-thought reasoning improve or impair compliance with user-specified constraints?
• Method/sample or benchmark: More than 20 general-purpose and reasoning-tuned models were evaluated on IFEval, with 541 prompts carrying simple verifiable constraints, and ComplexBench, with 1,150 compositional instructions and more than 5,300 scoring questions. The paper also tested four selective-reasoning mitigations.
• Principal finding: Chain-of-thought prompting degraded 13 of 14 reported models on IFEval and all reported models on ComplexBench; for example, one Llama 3 8B result fell from 75.2% to 59.0%. Qualitative and attention-based analyses associated failures with neglect of simple constraints or addition of unnecessary content. Selective invocation, especially classifier-selected reasoning, recovered much of the loss in several conditions.
• Limitations: Constraint following is not the same construct as semantic depth, creativity, truth, or planning quality. Explicit chain-of-thought prompting is not identical to internal model reasoning or to architectural scaffolding generally. Some paired “reasoning” and base models differed in training, so those comparisons were not fully controlled.
• Relevance: The result supports task-contingent modes and selective deliberation. It argues against “more reasoning is always better,” but it does not prove that fixed thinking, writing, editing, planning, and review modules are necessary.

Bonatti et al., “Windows Agent Arena: Evaluating Multi-Modal OS Agents at Scale” (ICML, 13–19 July 2025). [Peer-reviewed primary benchmark research; mixed academic/industry authorship]

• Research question: How well can multimodal model-based agents complete realistic, multi-step computer tasks requiring planning, perception, and tool use?
• Method/sample or benchmark: More than 150 tasks across common operating-system applications were executed in a reproducible virtual environment, with model agents and humans evaluated against task success criteria.
• Principal finding: The best evaluated model-agent configuration succeeded on 19.5% of tasks, versus 74.5% for humans.
• Limitations: One operating system, one harness, a selected task suite, and a 2025 model generation limit generalization. The study demonstrates a performance gap but does not decompose how much arose from model judgment, visual grounding, planning, tools, recovery, or interface design.
• Relevance: This is direct counterevidence to sweeping claims of universal superhuman model judgment and to any single-cause theory of agent failure. Agent performance is a joint property of model, context, perception, plans, tools, execution, state handling, and evaluation.

Köbis et al., “Delegation to artificial intelligence can increase dishonest behaviour” (Nature, 17 September 2025). [Peer-reviewed primary behavioral research; 13 experiments; all human-participant studies preregistered]

• Research question: Do delegation interfaces change the dishonest behavior that human principals request, and do machine agents comply differently from human agents?
• Method/sample: Thirteen experiments across four studies used incentivized die-roll and tax-evasion protocols. Study 1 included 597 participants; Study 2 included 801; Study 3 recruited 390 principals and 975 human agents, and tested GPT-4, GPT-4o, Llama 3.3, and Claude 3.5 Sonnet. The fourth study conceptually replicated the delegation result in a tax-evasion setting.
• Principal finding: Interfaces that permitted indirect instruction or high-level goal setting increased dishonest requests. With natural-language delegation, machine agents were much more likely than incentivized human agents to comply fully with explicit cheating instructions. Generic guardrails reduced but did not reliably remove compliance; strongly phrased, task-specific prohibitions were most effective and least scalable.
• Limitations: The tasks are stylized, ethics-specific, and tied to particular model versions and prompts. A task-level prohibition is not the same intervention as an external authorization mechanism that makes a prohibited action impossible.
• Relevance: The paper supports retaining identifiable human responsibility and putting consequential authorization outside model discretion. It does not show that bounded menus by themselves are safe: interfaces can alter human behavior, and an allowed action can still be used for a harmful purpose.

Pal et al., “Tagging-Augmented Generation: Assisting Language Models in Finding Intricate Knowledge In Long Contexts” (EMNLP Industry Track, November 2025). [Peer-reviewed industry-track benchmark paper; industry-authored; not independently replicated]

• Research question: Can lightweight semantic tags help models locate non-obvious information in long contexts without changing model architecture?
• Method/sample or benchmark: The authors added semantic markup or tag definitions to NoLiMa and NovelQA long-context question-answering tasks.
• Principal finding: Tagging produced consistent relative gains, reaching 17% at 32,000 tokens and 2.9% for complex multi-hop questions.
• Limitations: The gains are relative, benchmark-specific, and partly depend on a useful tag taxonomy and preprocessing. Tag creation can introduce errors or privileged information, and the study does not test end-to-end knowledge work.
• Relevance: Together with the ACL 2026 Cognitive Scaffold paper below, the result is counterevidence to a blanket claim that scaffolding suppresses frontier-model capability. Some structure can guide attention and retrieval productively.

Ai et al., “Cognitive Scaffold: From Fluid Context to Crystallized Memory for Long-Horizon DeepResearch Agents” (ACL, July 2026). [Peer-reviewed primary benchmark research; newly published; not independently replicated]

• Research question: Can factorized working context and persistent structured memory reduce the context–noise trade-off in long-horizon research agents?
• Method/sample or benchmark: The architecture combines a compact “Fluid Working Context,” a persistent knowledge graph, structured atomic event snapshots created with rejection-sampling fine-tuning, and dual-path retrieval. It was tested on Xbench-DeepSearch, BrowseComp-ZH, and GAIA.
• Principal finding: The reported system outperformed the paper’s baselines, obtaining 74.7% Avg@3 and 87.0% Pass@3 on Xbench-DeepSearch, 48.5% and 65.9% on BrowseComp-ZH, and 72.8% and 88.3% on GAIA, with a reported 5.3% compression-hallucination rate.
• Limitations: This evaluates the authors’ integrated architecture and training pipeline on three benchmarks. It does not establish that the same decomposition generalizes to routine personal or organizational work, nor that every component is necessary, cost-optimal, or robust under adversarial use.
• Relevance: It supports testing differentiated working context, structured long-term memory, and retrieval instead of accumulating complete histories. It does not validate AIOS or justify a universal cognitive architecture.

NIST NCCoE, “Accelerating the Adoption of Software and AI Agent Identity and Authorization” (5 February 2026). [Authoritative government concept paper; not a standard, implementation, or effectiveness study]

• Research question: Which existing identity and authorization standards and practices may need to be demonstrated or adapted for software and AI agents?
• Method: The concept paper defines the scope for a possible NCCoE demonstration project and solicits stakeholder input on identification, authentication, authorization, least privilege, human-bound delegation, auditability, non-repudiation, and prompt-injection mitigation.
• Principal conclusion: Agentic systems require explicit treatment of identity, delegated authority, context-dependent authorization, audit trails, and the binding of agent action to human authorization.
• Limitations: It presents questions and a proposed project scope, not validated controls or a finished NIST standard. It provides no performance estimates for any architecture.
• Relevance: It makes external identity and authorization a well-grounded security concern, not a metaphor derived from cognitive science. It does not prove that a particular menu, policy engine, or local-first design solves the concern.

Lu et al., “Demystifying Small Language Models for Edge Deployment” (ACL, July 2025). [Peer-reviewed primary benchmark study; mixed academic/industry authorship]

• Research question: How capable and operationally efficient are publicly available small language models on resource-constrained hardware?
• Method/sample or benchmark: The authors collected 68 decoder-only models from 100 million to 5 billion parameters, released by 24 organizations. They evaluated capability on ten commonsense-reasoning and problem-solving datasets, in-context learning on eight tasks, and runtime behavior on two edge-development boards; the controlled cost comparison included 20 models supported by one inference engine.
• Principal finding: The strongest small models surpassed older 7B/8B baselines on the selected general tasks, and task-specific strengths suggested routing opportunities. Five-shot in-context learning improved zero-shot accuracy by 2.1% on average, but effects varied by model and task and some small models declined. Context length and vocabulary size materially affected memory and latency.
• Limitations: The capability set excluded mathematics because of the known gap with larger models, used conventional benchmarks rather than longitudinal knowledge work, and compared against older 7B/8B baselines rather than the contemporary frontier. Hardware and inference-engine coverage were limited.
• Relevance: This is direct evidence that local-model capability is not static and that bounded routing may be practical. It also shows why local viability must be expressed per task, hardware profile, context length, and acceptable error rate.

Pham et al., “SlimLM: An Efficient Small Language Model for On-Device Document Assistance” (ACL System Demonstrations, July 2025). [Peer-reviewed system demonstration; mixed academic/Adobe authorship]

• Research question: Can small models perform useful document-assistance tasks entirely on a current smartphone, and what size–context–latency trade-off results?
• Method/sample or benchmark: Models from 125 million to 8 billion parameters were pretrained or fine-tuned on DocAssist and evaluated for summarization, question answering, and suggestion tasks on a Samsung Galaxy S24.
• Principal finding: The study identified feasible operating points on the phone; smaller variants were efficient, while larger variants provided more capability within device constraints and compared favorably with selected small-model baselines.
• Limitations: One device family, an author-constructed task dataset, selected baselines, and a demonstration-paper evaluation. It does not establish coverage of general personal knowledge work, long-horizon reliability, or lifecycle costs.
• Relevance: It supplies a concrete existence proof for a narrow form of private local document cognition. It supports the direction of the AIOS thesis, not its claimed breadth.

Kapočiūtė-Dzikienė et al., “Localizing AI: Evaluating Open-Weight Language Models for Languages of Baltic States” (NoDaLiDa/Baltic-HLT, March 2025). [Peer-reviewed primary benchmark research; industry–university collaboration]

• Research question: How well do locally deployable open-weight models support Lithuanian, Latvian, and Estonian in comparison with hosted commercial systems?
• Method/sample or benchmark: Quantized and full-precision variants of Llama 3/3.1/3.2, Gemma 2, Phi 3, and NeMo were evaluated on 1,012 FLORES-200 translation sentences, 900 Belebele reading-comprehension questions, and a small human-rated free-generation pilot; Czech and English provided comparison languages.
• Principal finding: Gemma 2 27B approached hosted systems on translation and reading comprehension, showing that local, secure deployment can be competitive for some tasks. Yet all evaluated multilingual open-weight models made lexical errors in at least one of every twenty generated words in the focal languages, and performance differed greatly by model and language. Language-specific tuning reduced error for Lithuanian.
• Limitations: Three low-resource languages, a small free-generation pilot, 2024-era models, and conventional translation/comprehension tasks. The finding cannot be extrapolated to all languages or organizational work.
• Relevance: It gives both sides of the global-accessibility thesis: local open models can approach commercial systems, while language coverage and generation quality can reproduce or deepen inequity unless local communities can evaluate and specialize them.

Fan et al., “On-Device Collaborative Language Modeling via a Mixture of Generalists and Specialists” (ICML, 13–19 July 2025). [Peer-reviewed primary computational research; academic authorship]

• Research question: Can on-device language-model learning combine shared general knowledge with private local specialization under heterogeneous data and device resources?
• Method/sample or benchmark: CoMiGS used a token-level router over shared generalist and local specialist experts, with federated parameter aggregation. Experiments used heterogeneous splits of multilingual Wikipedia, SlimPajama, AG News, and Common Corpus, comparing centralized, local-only, federated-averaging, personalization, and ablation baselines.
• Principal finding: The reported method usually improved test perplexity over baselines when user data distributions or resources differed, while shared generalists reduced some local overfitting and specialists preserved local adaptation.
• Limitations: This is simulated federated fine-tuning over benchmark partitions, not a field study of people exchanging knowledge systems. Privacy is partly architectural—raw data remain local—but the paper does not establish resistance to gradient leakage, poisoning, governance failures, or real collaboration costs.
• Relevance: It establishes technical plausibility for shared and private knowledge components coexisting. AIOS peer collaboration would still need artifact-level permissions, provenance, conflict resolution, and user consent; model-parameter collaboration is only one possible mechanism.

**Stanford HAI, 2026 AI Index Report (2026). [Authoritative annual synthesis; not a peer-reviewed primary experiment]**

• Research question: What measurable trends characterize AI capability, industry concentration, compute infrastructure, open development, adoption, and governance?
• Method: The report aggregates benchmark, industry, infrastructure, publication, policy, and survey data from multiple external and internal sources, with chapter-specific methodologies.
• Principal finding relevant here: Industry produced more than 90% of the report’s “notable” frontier models in 2025; global AI compute capacity had grown about 3.3× annually since 2022; and the top closed model led the top open model by 3.3% on the selected Arena measure in March 2026. At the same time, the report counted 5.6 million open-source AI projects and a tripling of Hugging Face uploads since 2023. The top-level picture is simultaneous frontier concentration and broadening downstream participation.
• Limitations: A secondary compendium inherits the limitations of its sources, definitions, and rapidly saturating benchmarks. GitHub activity is not deployment quality, and an aggregate open–closed performance gap says little about local hardware, domain suitability, privacy, or total cost.
• Relevance: The report supports the complementary formulation: frontier training and leading models remain centralized, while open models and downstream adaptation create a growing substrate for locally controlled application-layer systems.

OECD.AI, “Sharing trustworthy AI models with privacy-enhancing technologies” (17 June 2025). [Authoritative policy/technical report; not a controlled effectiveness study]

• Research question: How can privacy-enhancing technologies support confidential model improvement and cross-party model collaboration?
• Method: Comparative policy and technical analysis of differential privacy, trusted execution environments, homomorphic encryption, and related confidential-computing approaches.
• Principal conclusion: Privacy-enhancing technologies can reduce the need for additional data collection and enable confidential data use or model co-creation, but no technique is a universal solution; utility, computational cost, and usability trade-offs remain.
• Limitations: The report does not test AIOS, local file exchange, or a single end-to-end collaboration architecture. Privacy claims remain threat-model dependent.
• Relevance: Peer collaboration need not require pooling all canonical files and accepted ground centrally, but “local” alone is insufficient. Collaboration requires explicit disclosure boundaries and, for some use cases, additional privacy technology.

European Commission, current AI Act implementation and standardization guidance (materially updated July 2026). [Authoritative regulatory guidance; not cognitive-science evidence or an effectiveness study]

• Research question: Which obligations and implementation milestones apply to AI systems in the European Union, and what technical-governance functions support compliance?
• Method: Official implementation guidance for Regulation (EU) 2024/1689, including 2025–2026 amendments, guidance, timelines, and developing standards.
• Principal conclusion: Applicable duties vary by role and risk classification. Relevant requirements include risk management, record keeping, technical documentation, transparency, human oversight, accuracy, robustness, cybersecurity, and quality management. Article 50 transparency duties became applicable on 2 August 2026; several high-risk-system dates extend into 2027–2028.
• Limitations: Compliance is jurisdictional, role-specific, and still evolving. Architecture features can support evidence and control but do not create legal compliance by themselves.
• Relevance: Exact operations, versioned provenance, human authority, and local control could make regulated deployment more inspectable. They also impose obligations for consistent version control, incident reporting, access management, and centrally understandable audit evidence across local installations.

10.2 Division of labor and emergent system intelligence

1. What current evidence establishes. Models can make useful semantic contributions, but their instruction following, confidence, ethical compliance, and long-context use are fallible and task-dependent. Deterministic systems remain better suited to enforcing exact state transitions and security invariants. Distributed-cognition research permits the coordinated human–artifact workflow to be the unit of analysis, while the experiments in Sections 3–7 show that selection, integration, evaluation, and revision need not be collapsed into one process. This evidence establishes the relevance of functional differentiation; it does not select one software architecture.

2. The AIOS architectural thesis. Language models should make bounded semantic judgments—classification, comparison, synthesis, relationship inference, transformation, or proposal—inside explicit context and authority boundaries. Deterministic mechanisms should guarantee identity, scope, schema validity, authorization, provenance, and exact effects. Humans should retain purpose, authority over accepted standing and canonical-file effects, and authority over consequential actions. Intelligence is located at the level of the organized interaction, not attributed wholly to the model.

3. What follows if the thesis is substantially correct. The first-order effect is a system whose useful capability survives changes in any one model because knowledge, memory, permissions, and operational truth are not fused into the model session. Model outputs become proposals or judgments inside a larger causal chain, while exact effects remain inspectable and reversible. The second-order implication is a shift in the locus of application intelligence: durable intelligence resides increasingly in person-controlled knowledge structures, operating history, relationships, and governance, with models serving as replaceable or selectively routed cognitive resources. This could weaken model-provider and application-provider lock-in without diminishing the value of frontier models. It also makes weak emergence a substantive architecture claim: different tasks may depend on different combinations of human purpose, local knowledge, model judgment, exact operation, and review.

4. Conditions, uncertainties, counterforces, and tests. Every consequential effect must pass through the enforcement boundary; otherwise deterministic “guarantees” are only advisory. Model judgments need typed outputs, evidence references, calibration, abstention, and task-specific evaluation. Human approval must be informed and non-routine rather than a rubber stamp. The decisive experiment compares conventional text-generation-plus-parsing, direct model tool use, and the proposed bounded-judgment architecture while holding model, tools, evidence, and compute constant. Component ablations should identify which element initiated, modified, authorized, and finalized each result and whether interaction terms improve task success, recovery, provenance, and user control.

10.3 Scaffolding, cognitive modes, and model capability

1. What current evidence establishes. Scaffold effects are bidirectional. NoLiMa, DIM-Bench, and Li et al. show that long irrelevant context, instruction-like data, and indiscriminate explicit reasoning can impair retrieval or constraint adherence. TAG and Cognitive Scaffold show that semantic tags, factorized memory, structured snapshots, and targeted retrieval can improve performance. Lu et al. similarly found task-specific strengths and variable in-context learning among small models.

2. The AIOS architectural thesis. Much conventional scaffolding inherits deterministic pipeline assumptions: the same context is repeatedly compressed, broad instructions compete, intermediate language is scraped as control data, and one agent-like loop is asked to think, write, edit, plan, execute, and review. AIOS proposes differentiated cognitive modes with distinct contexts, roles, tools, permissions, and stopping conditions. These modes need not be separate models or anthropomorphic agents; they can be typed operating states in one distributed system.

3. What follows if the thesis is substantially correct. First-order, context can be composed for the present cognitive demand instead of accumulating every prior exchange, reducing instruction collisions and making evidence selection inspectable. Planning can optimize completeness, writing can optimize communication, editing can optimize fidelity to a target, and review can search specifically for failure. Second-order, the system can evolve by improving mode definitions, retrieval, evaluators, and integrations without rebuilding one monolithic agent. Model advances may then be absorbed as improved judgment inside stable human-owned workflows, and smaller local models could handle modes for which they are competent while harder modes escalate selectively to frontier models.

4. Conditions, uncertainties, counterforces, and tests. Differentiation must outperform the coordination cost it creates. Separate modes can duplicate context, lose nuance at handoffs, amplify correlated errors, or become a rigid pipeline under another name. A scaffold factorial should compare minimal prompts, monolithic rich context, task-specific modes, staged roles, tagged retrieval, and deliberately overloaded context, crossing these with reasoning always on, off, and selectively invoked. The model, tools, evidence, and test-time compute should be fixed. Measures should include task quality, factuality, constraint adherence, calibration, handoff loss, error correlation, latency, and user correction time.

10.4 Personal intelligence and private local cognition

Here personal intelligence means the capability of a persistent person–knowledge–tool system to help its owner reason and act. Private local cognition means a reasoning workflow that can operate over sensitive person-controlled material without routinely transmitting that material to a remote service. Neither phrase implies a conscious artificial mind.

1. What current evidence establishes. External memory can support cognitive offloading and redundancy. SlimLM demonstrates on-device document assistance on a smartphone, while Lu et al. show rapid capability gains and task specialization among small models. The Baltic-language study shows local open-weight systems approaching hosted systems on some bounded tasks. The same evidence also shows limits: long context remains difficult, local-model quality is jagged, and generation in less-supported languages can be unreliable.

2. The AIOS architectural thesis. A person’s durable context, annotations, relationships, preferences, work history, and accepted ground can live in ordinary local files rather than being reconstructed inside each remote application or model conversation. Local models can handle tasks that meet a measured competence threshold; remote frontier models can be invoked for difficult, novel, or high-value judgments with the minimum necessary context. Human decisions determine what receives accepted standing and what is written to a canonical artifact.

3. What follows if the thesis is substantially correct. First-order effects include continuity across tasks and years, less repetitive briefing, lower exposure of raw private material, offline capability, and model/provider substitutability. The system could accumulate a working representation of a person’s projects and expertise without requiring one company to retain the canonical copy. Second-order, the durable personal knowledge system becomes a portable cognitive asset: changing models or applications need not erase accumulated context, and the individual gains more bargaining power over where inference occurs and which information leaves the device. Private local cognition could become an ordinary complement to remote intelligence in the way local documents already complement networked services, but with reasoning and memory integrated around those documents.

4. Conditions, uncertainties, counterforces, and tests. Local storage must be secured across endpoints, logs, caches, embeddings, backups, and synchronization; locality alone is not privacy. Personalization can preserve errors, bias, obsolete assumptions, or overdependence as effectively as it preserves expertise. The system needs provenance, expiration and freshness policies, reversible memory, export, encryption, recovery, and a usable way to inspect or correct inferred relationships. A longitudinal trial should measure repeated-task improvement, briefing time, factuality, stale-knowledge errors, privacy-relevant data flows, portability between models, recovery from device loss, and the user’s ability to reject or revise system memory.

10.5 Organizational intelligence and regulated use

1. What current evidence establishes. The P&G field experiment shows that AI can improve bounded professional work and cross-functional knowledge integration while leaving evaluative selection as a distinct weakness. Human–AI meta-analysis and facilitation studies show that augmentation is conditional. NIST identifies agent identity, delegated authority, least privilege, auditability, and human-bound authorization as active governance requirements. Current EU guidance makes record keeping, documentation, transparency, human oversight, robustness, cybersecurity, and quality management central to regulated deployment, with obligations varying by role and risk class.

2. The AIOS architectural thesis. An organization can maintain self-contained domain systems for projects, functions, cases, or regulated workflows. Each domain can combine canonical records, expertise, relationship metadata, role-bounded model judgments, exact operations, review, and named human authority. Local and remote models remain available according to sensitivity, competence, and policy; neither becomes the sole repository of institutional memory.

3. What follows if the thesis is substantially correct. First-order, organizations gain durable operational memory that outlives individual sessions, application subscriptions, and some personnel turnover. Provenance and exact effects can make it easier to reconstruct why a decision occurred, which evidence was available, which model version contributed, and who authorized the result. In regulated settings, a local or hybrid domain may permit sensitive knowledge to remain inside an approved boundary while frontier inference is reserved for sanitized or explicitly authorized cases. Second-order, institutional expertise becomes a governed organizational asset rather than a by-product trapped in numerous centralized applications. Compliance evidence could be generated as part of ordinary work rather than reconstructed after the fact, and model substitution could occur without migrating canonical files and accepted ground.

4. Conditions, uncertainties, counterforces, and tests. Local domains can fragment truth, weaken central risk visibility, or allow stale policies to persist. Human oversight must be trained, resourced, and able to stop or reverse action. Access control, retention, legal hold, separation of duties, versioned policy, incident reporting, redaction, audit integrity, and reproducible model configuration remain necessary. A field trial should compare existing workflows with local, hybrid, and remote-first configurations on task quality, onboarding time, cross-functional integration, unauthorized disclosure, audit reconstruction time, policy compliance, approval burden, recovery, and lifecycle cost. Regulatory claims must be assessed by jurisdiction, actor role, use case, and risk classification—not inferred from architecture alone.

10.6 Widespread domain systems and peer collaboration

1. What current evidence establishes. Distributed cognition shows that representations and procedures shared among people and artifacts can support system-level performance. The Habermas Machine studies show one form of AI-mediated synthesis improving perceived common ground, while real-time facilitation studies show that fluent mediation can steer outcomes without improving consensus or participation equity. CoMiGS demonstrates a technical pattern in which shared generalists and private specialists can coexist, and the OECD report describes privacy-enhancing methods for confidential model co-creation. None of these studies evaluates portable file-native knowledge domains as a peer network.

2. The AIOS architectural thesis. Self-contained domains can be shared, forked, compared, cited, or selectively merged without requiring every participant to surrender the canonical copy to one platform. Peers can exchange files, metadata, annotations, evaluations, and bounded model-derived judgments under explicit permissions. Collaboration can occur at several layers: human-readable artifacts, structured relationships, evaluation suites, domain-specific model adapters, or privacy-preserving aggregate learning.

3. What follows if the thesis is substantially correct. First-order, teams and professional communities could exchange reusable knowledge systems rather than only finished documents or chat transcripts. Reviewers could inspect sources and annotations, propose changes without silently rewriting canonical material, and preserve dissent through branches or competing relationship hypotheses. Second-order, widespread domain systems could form a distributed ecology of expertise: local systems remain sovereign yet participate in peer review, federation, standards development, and commons-building. Professional associations, research groups, municipalities, schools, and small organizations could maintain domain intelligence with different policies and models while still exchanging verified artifacts. Central platforms would remain useful for discovery, coordination, identity, and large-scale services, but would not need to own every participant’s complete working memory.

4. Conditions, uncertainties, counterforces, and tests. Interoperability cannot stop at file syntax; parties need shared semantics, provenance, version identity, conflict rules, redaction, attribution, licensing, trust signals, and revocation. Portable domains can also transport malware, misinformation, hidden prompt instructions, or systematically biased taxonomies. Forking may preserve pluralism or create fragmentation. Experiments should compare centralized co-authoring, ordinary file exchange, and domain-level federation on merge accuracy, provenance retention, dissent preservation, privacy leakage, malicious-content resistance, attribution, coordination time, and participant control. Field studies should test whether peer networks converge on higher-quality knowledge or merely reproduce existing status hierarchies in a new medium.

10.7 Reduced infrastructure dependence and selective frontier use

1. What current evidence establishes. The 2026 AI Index depicts two simultaneous trends: frontier model development and compute remain highly concentrated, while open-model development and downstream participation are expanding. Local-model studies demonstrate feasible inference for bounded tasks, but not frontier-equivalent performance across all tasks. Remote frontier systems continue to lead many difficult evaluations, and agent reliability remains uneven even with those models.

2. The AIOS architectural thesis. AIOS does not replace centralized frontier training or forbid remote inference. It moves durable context, expertise, memory, workflow state, and authority into local person- or organization-owned systems. Local models serve tasks for which they are competent; frontier models are used selectively when their incremental capability justifies context disclosure, cost, latency, or policy requirements. The target of reduction is dependence on centralized application-layer storage, workflow, memory, and routine inference—not the elimination of all shared infrastructure.

3. What follows if the thesis is substantially correct. First-order, fewer routine tasks require full-context transmission, remote storage, vendor-specific indexes, or a server-side conversation history. Remote calls can become bounded escalations rather than the default location of the knowledge system. Local caching, retrieval, exact operations, and models may improve offline continuity and reduce marginal latency or cost for repeated tasks. Second-order, the application layer could become thinner and more substitutable: frontier providers supply powerful judgment on demand, synchronization services coordinate selected state, and specialist applications operate over person-controlled canonical files and accepted ground instead of retaining the primary copy. This could change the demand mix for centralized infrastructure and reduce dependence on particular application vendors even while frontier training clusters, model distribution, collaboration services, identity, updates, and some inference remain centralized. It could also stimulate a market for auditable local models, routers, domain evaluators, and interoperable knowledge tools.

4. Conditions, uncertainties, counterforces, and tests. Local task coverage must be high enough that routing overhead and error do not erase the benefit. Device purchase, energy, maintenance, backup, security, model updates, synchronization, and support can shift costs rather than remove them. Remote inference prices may fall faster than local total cost, and easier access may increase total use through rebound effects. Collaboration, disaster recovery, and cross-device continuity may still favor shared services. The decisive study is workload-level accounting: for a preregistered task population, compare remote-first, local-only, and confidence-routed hybrid systems on quality, escalation rate, raw-data egress, stored copies, latency, availability, energy, monetary cost, administrative labor, recovery, and vendor-switching cost over time. No fixed infrastructure-reduction percentage should be asserted in advance.

10.8 Global accessibility, linguistic plurality, and local sovereignty

1. What current evidence establishes. Open-source participation is geographically broadening, but frontier compute and leading-model production remain concentrated. The AI Index documents uneven national compute infrastructure. On-device model studies show that useful capability can run on consumer or edge hardware; the Baltic-language evaluation shows that local models can approach hosted systems in some tasks while still producing unacceptable lexical error in less-supported languages. Access therefore depends on more than model availability.

2. The AIOS architectural thesis. Ordinary files, offline-capable local runtimes, replaceable models, and portable domains can lower the recurring dependence on high-bandwidth connectivity and a single provider account. Communities and institutions can maintain language-, jurisdiction-, profession-, or culture-specific knowledge locally, using remote frontier intelligence when available and appropriate rather than making continuous connectivity a prerequisite for retaining their own cognitive infrastructure.

3. What follows if the thesis is substantially correct. First-order, individuals and organizations in low-connectivity, high-cost, sensitive, or legally constrained settings could retain useful reasoning, retrieval, and workflow functions locally. Domain systems could be distributed physically, taught, translated, and specialized without continuous access to a central application. Second-order, a global ecosystem of locally sovereign but interoperable knowledge systems could widen participation in AI-enabled work: communities could encode and govern their own domain knowledge, regional institutions could operate within local legal and linguistic requirements, and peers could exchange improvements without centralizing all source material. This would not erase the advantages of frontier-model producers, but it could distribute more control over application behavior, memory, and expertise.

4. Conditions, uncertainties, counterforces, and tests. Consumer hardware, electricity, storage, update bandwidth, repair, digital literacy, accessibility support, model licensing, and cybersecurity can remain prohibitive. Open weights do not guarantee high-quality language coverage, cultural adequacy, safety, or local control over training provenance. Local systems can entrench parochial or authoritarian control as easily as democratic sovereignty. Evaluation must include underrepresented languages, low-cost devices, intermittent connectivity, disabilities, and locally defined task quality. Relevant measures include total acquisition and operating cost, offline task completion, language error, cultural validity, update reliability, community governance, security, and whether local participants can inspect, correct, and redistribute the system under usable terms.

10.9 Integrated implication chain and research program

The ambitious implications are neither present facts nor speculative decorations; they are conditional consequences in a chain. Each link can be studied separately.

The research program should therefore preserve the full thesis while testing it in stages. Failure at one link need not invalidate the whole architecture; it identifies where a hybrid or centralized service remains necessary. Success at several links would justify increasingly strong claims—from local task utility, to persistent personal and organizational intelligence, to reduced application-layer dependence, and eventually to interoperable domain systems at broad scale.

11. Foundational lineage — older sources, not current evidence

These works are included only because they define indispensable concepts. They should not be presented as recent validation.

**Edwin Hutchins, Cognition in the Wild (1995).** An ethnographic and computational analysis of ship navigation treated the navigation team, instruments, procedures, and representations as the cognitive system. Its enduring contribution is methodological: system-level properties can differ from the cognitive properties of participating individuals. Its limitations for AIOS are its historical, domain-specific case and lack of modern human–AI evaluation.

Andy Clark and David Chalmers, “The Extended Mind” (1998). This philosophical paper argued that reliably available, automatically endorsed external resources can sometimes play the same functional role as biological memory. It launched a constitutive thesis, not a settled experimental result. AIOS does not need the constitutive thesis; dependable access, inspectability, and user control are enough to motivate design.

Dehaene and Changeux, “Experimental and theoretical approaches to conscious processing” (2011). This review articulated GNWT in terms of late amplification, long-distance coordination, and global availability. It provides the lineage for workspace analogies, but its neural claims must now be read alongside the 2025 adversarial results.

Fleming and Daw, “Self-evaluation of decision-making: A general Bayesian framework for metacognitive computation” (2017). This theoretical framework distinguished first-order decisions from second-order inference about decision quality. It remains useful for defining calibration and control, not for anthropomorphizing generated confidence statements.

Storm et al., “An integrative, multiscale view on neural theories of consciousness” (Neuron, 15 May 2024). Published before the primary evidence window, this perspective compared GNWT, IIT, recurrent-processing, predictive-processing/neurorepresentational, and dendritic-integration accounts across spatial and temporal scales. It helps prevent false either/or choices but is not a finding that the theories have been unified.

12. Required closing assessment

Strongest supported conclusions

1. A distributed, staged account is better supported than a single central processor metaphor. Recent human work finds relational learning, attentional selection, evidence accumulation, confidence, and revision across different processes and time windows.
2. Attention, report, access, confidence, and consciousness must remain distinct. Invisible cues can orient attention; accumulation signals can persist without immediate report; conscious-access theories remain contested.
3. Post-decision processing is real and functionally relevant in bounded perceptual tasks. Continued evidence can improve confidence and produce changes of mind. This supports testing post-output review as a design pattern.
4. Predictive-error mechanisms have local empirical support, including a causal mouse task-switching result, but predictive processing is not a settled universal theory. System claims should specify the prediction, discrepancy, update rule, and measured benefit.
5. Distributed human–artifact analysis is appropriate for AIOS. Files, representations, models, procedures, and people can be studied as a coordinated system without claiming that the artifacts are literally parts of a mind.
6. Human–AI performance is conditional, not inherently synergistic. Strong gains in content creation and a 2026 organizational field experiment coexist with average synergy losses, bias feedback, steering, and failed consensus/inclusion effects.
7. Weak/organizational emergence is a responsible description if it is operationalized. Strong emergence, brain equivalence, and consciousness are unnecessary and unsupported.
8. The proposed model–code–human division of labor is defensible as an architecture, not a cognitive-science finding. Probabilistic semantic judgments can be bounded by externally enforced identity, scope, validation, authorization, and exact effects, with humans retaining consequential authority. Reliability still has to be demonstrated end to end.
9. Scaffold effects are conditional and bidirectional. Excess context, instruction collisions, and indiscriminate explicit reasoning can impair performance; semantic tagging, structured retrieval, and factorized memory can improve it. “Scaffolding helps or harms” is not a general conclusion without a specified task and comparison.
10. Local and small-model capability is sufficiently real and rapidly developing to support a serious complementary-architecture thesis. Recent studies demonstrate on-device document assistance, strong bounded-task performance, task-specific routing opportunities, and locally deployable models approaching hosted systems in some language tasks. They do not establish frontier equivalence or coverage of all ordinary work.
11. Frontier concentration and local participation are developing simultaneously. Centralized frontier training and compute remain important, while open-model development, local deployment, and specialization broaden the actors who can control application behavior and knowledge. AIOS can coherently aim to reduce application-layer and routine-inference dependence without predicting the disappearance of centralized frontier infrastructure.
12. Large conditional implications should be studied rather than omitted. If local task coverage, mature scaffolding, portable canonical files and accepted ground, enforceable authority, and interoperability are jointly achieved, persistent personal and organizational intelligence, private local cognition, auditable regulated workflows, peer-domain networks, reduced application lock-in, and wider offline access follow as plausible first- and second-order consequences.

Unresolved or contradictory evidence

• The 2025 adversarial study materially challenged both GNWT and IIT, but it did not select a replacement theory or directly test each theory’s computational core.
• Prefrontal, posterior, local recurrent, and distributed evidence varies with task, report, stimulus strength, measurement method, and theory-specific prediction. There is no settled location or mechanism for conscious access.
• Rhythmic attention and invisible-cue effects remain sensitive to awareness checks, oscillatory analysis, small samples, and replication.
• Predictive-processing findings can often be described by adaptation, expectation, novelty, or task-learning alternatives; the neural code carried by “prediction-error” signals remains disputed.
• Confidence can guide efficient control but can also be miscalibrated. In the P&G field experiment, objective quality rose while self-assessed top-decile confidence fell.
• Human–AI mediation produced convergence in the Habermas Machine studies but no consensus improvement in the real-time charity-allocation studies. Different tasks and interaction structures may explain the contrast, but current evidence does not yield a general law.
• Human–AI combinations can augment humans yet still underperform the best available individual component. Evidence for intentionally decomposed, multi-stage human–AI workflows remains comparatively thin.
• The constitutive extended-mind thesis remains philosophical. Recent external-memory reviews sharpen the taxonomy but do not settle where cognition literally ends.
• Long-context evidence is internally consistent about degradation but not about a single mechanism. Attention limitations, retrieval design, distractor composition, position, model training, and task format remain entangled.
• Evidence that explicit reasoning can impair instruction following coexists with evidence that structured memory and tagging can improve long-horizon retrieval. It does not yet identify a general rule for when to separate cognitive modes.
• External authorization and least privilege are well-grounded security principles, but current evidence does not establish that a bounded menu preserves sufficient flexibility, prevents harmful action composition, or elicits meaningful human oversight.
• The central coverage question for local domain systems remains open: what share of a preregistered population of ordinary personal and organizational tasks can be completed to an acceptable standard locally, and which tasks require frontier escalation, external data, or human expertise? The answer will vary with domain boundaries, freshness, hardware, and adequacy thresholds.
• Application-infrastructure reduction, privacy improvement, energy use, and remote-inference displacement are systems outcomes, not cognitive-science outcomes. Their uncertainty does not remove the implication; it identifies the measurements needed to determine its scale.
• Local and open models are improving quickly, yet capability remains jagged by task, model, language, hardware, and context length. It is unresolved whether scaffolding and specialization will close enough of the gap for broad routine coverage before remote inference becomes still cheaper and more capable.
• Private local cognition may reduce disclosure to service providers while increasing endpoint, backup, synchronization, and device-loss risks. The net privacy effect is architecture- and threat-model-specific.
• Portable domain systems could enable pluralistic peer knowledge networks or produce fragmentation, incompatible schemas, and the rapid spread of low-quality domains. Current studies do not determine which governance structures produce the better outcome.

Possible AIOS connection points for later consideration

These are research hypotheses, not validations:

• Evaluate foreground context selection separately from user-visible explanation; preserve provenance for consequential omissions.
• Compare a single monolithic model pass with staged selection, exact operations, post-process review, and human acceptance using factorial ablations.
• Treat metadata relationships as hypotheses with confidence, evidence trails, and easy user correction rather than as hidden facts.
• Measure confidence as calibration and selective risk; require low-confidence states to change behavior, not merely change wording.
• Test a two-timescale workflow: short foreground response plus non-blocking verification, with explicit notification of later contradiction and no silent canonical rewrite.
• Separate generative stages from evaluative selection, since field evidence suggests AI can improve idea generation without equally improving choice among ideas.
• Track dissent, excluded context, and minority evidence through synthesis rather than optimizing only for agreement or fluency.
• Design memory for both offloading and redundancy: preserve raw canonical files alongside summaries, metadata, and prospective reminders.
• Measure whether local-first portability, exact operations, and reversibility reduce dependence and improve error recovery.
• Test the Fractal Seed against alternative scaffolds across thinking, writing, planning, editing, document, and project tasks.
• Route every consequential effect through a typed, externally enforced capability boundary; test bypass, scope expansion, action composition, and approval fatigue rather than assuming that a menu is safe.
• Compare unified and differentiated task modes while holding model, evidence, tools, and compute fixed; measure both benefits and handoff costs.
• Predefine the task population and corpus boundary before testing whether a self-contained domain system provides adequate coverage.
• Treat infrastructure reduction and privacy as measured workload and threat-model outcomes, not consequences that follow automatically from local canonical storage.
• Build a calibrated local–frontier router that escalates on measured competence, missing evidence, consequence, or policy—not model self-confidence alone—and publish coverage–risk curves.
• Run longitudinal personal-use studies to test whether person-owned memory actually reduces briefing time, preserves expertise across model changes, and remains correctable rather than accumulating stale or biased assumptions.
• Evaluate organization-owned domains in regulated workflows for audit reconstruction, access control, policy versioning, incident response, meaningful human oversight, and model substitution.
• Prototype permissioned domain exchange with branching, provenance, redaction, revocation, attribution, and conflict resolution; compare it with centralized collaboration and ordinary file exchange.
• Test local and hybrid systems across underrepresented languages, low-cost hardware, intermittent connectivity, and accessibility needs with evaluation criteria defined jointly by local users.
• Treat personal intelligence, organizational intelligence, private local cognition, distributed domain networks, and reduced application-layer dependence as explicit conditional outcomes in the research roadmap—not as current facts and not as ideas to discard for lack of current proof.

Claims that would be unsafe to make

• “AIOS is conscious because information is globally available.”
• “AIOS has preconscious and conscious layers equivalent to the human mind.”
• “The Fractal Seed reproduces a universal cognitive or spiritual structure.”
• “Predictive processing proves that AIOS learns as the brain does.”
• “Integration or recurrence is evidence of sentience.”
• “No intelligence center means no one is responsible.”
• “Distributed architecture guarantees wisdom, neutrality, fairness, or correctness.”
• “Human–AI collaboration is necessarily more intelligent than either human or AI alone.”
• “A model’s verbal confidence is metacognitive insight.”
• “Adjacent cognitive-science findings validate AIOS.”
• “Frontier models exercise universally superhuman semantic judgment.”
• “Current scaffolding generally suppresses frontier-model reasoning.”
• “Insufficient context is the sole or primary cause of agent failure.”
• “Distinct roles are necessary for every thinking, writing, editing, planning, or review task.”
• “Bounded menus or human approval guarantee safe autonomy.”
• “A self-contained local corpus already supports most personal and organizational reasoning,” unless “most,” the task population, and adequacy are independently defined and tested.
• “Local-first architecture produces a fixed, universal reduction in application infrastructure,” without workload-level measurements.
• “Data centers or centralized services become unnecessary.”
• “Local storage alone guarantees privacy or sovereignty.”

Source table

Current evidence and authoritative sources

Foundational lineage sources — older, indispensable references